High-precision pressure sensor compensation method based on deep learning

Abstract

The invention discloses a high-precision pressure sensor compensation method based on deep learning, and the method comprises the steps: forming training data based on a low-precision pressure sensor, a high-noise pressure and temperature output curve of a high-precision pressure sensor, and a standard pressure and temperature output curve in combination with an image of a calibrated pressure and temperature output curve during delivery; and training the generative adversarial network. And based on the trained generative adversarial network, generating an image of the preliminarily compensated pressure and temperature output curve of the to-be-tested pressure sensor. Then, at a user, training a neural network based on historical data actually acquired by a to-be-detected pressure sensor and the preliminarily compensated pressure detection value at the corresponding temperature to obtain a compensation model of the pressure sensor; and finally, inputting the actually tested detection value of the pressure sensor and the temperature parameter into the compensation model to obtain an actually compensated pressure detection value. According to the method, when a user trains the neural network to obtain an accurate compensation model of the pressure sensor, self-calibration of pressure is omitted, the model can be rapidly trained, the capacity of training data can be reduced, and good practicability is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of pressure sensor compensation, and specifically relates to a high-precision pressure sensor compensation method based on deep learning. Background Art

[0002] In digital warehouse management, it is necessary to deepen the management of inventory materials, give full play to the advantages of overall coordination and allocation of inventory materials, improve the warehouse-warehouse coordination mechanism, build intelligent and unmanned warehouse management functions, and efficiently guarantee the supply of power grid materials. At present, by adopting information-based supervision and management methods, combined with RFID, face recognition and big data analysis technologies to achieve security monitoring, supervision and inquiry, and provide digital decision-making support, the intelligent operation of material management is promoted to move forward, and the digital management level of enterprise professional warehouses is comprehensively improved. Specifically, the first is to use the Internet of Things and data analysis technology, combined with intelligent weighing shelves, face recognition and other means to realize the automation and intelligence of spare parts and construction tools (safety tools) management in the warehouse; the second is to combine sensor technology and RFID sensing chips to monitor and manage the status and location of items in real time; the third is to use big data analysis intelligent algorithms to deeply optimize warehouse operation data for analysis and statistics.

[0003] The sensing technology refers to the technology of sensing and collecting physical quantities using various sensors. In the smart logistics warehouse, various types of sensors can be used to monitor the status and location information of items, such as temperature sensors, humidity sensors, acceleration sensors, pressure sensors, etc. By connecting these sensors to the computer system, real-time monitoring and management of items can be achieved. Through various sensors, automatic identification technology, automated equipment, etc., the warehouse's goods warehousing, outbound, inventory, sorting and other operations are automated, thereby reducing manual intervention and improving warehouse operation efficiency and accuracy.

[0004] The intelligent weighing shelf is a shelf that can automatically identify and record the weight of goods. It can realize real-time monitoring and management of the weight of goods by installing pressure sensors, photoelectric sensors and other equipment. This shelf can greatly improve the efficiency and accuracy of warehouse management and reduce the error rate and cost of manual operation. However, the test accuracy of the pressure sensor is easily affected by the accuracy of the sensor itself and the ambient temperature. Therefore, the pressure sensor needs to be compensated. Summary of the invention

[0005] The purpose of the present invention is to provide a high-precision pressure sensor compensation method based on deep learning, aiming to solve the above-mentioned problems.

[0006] The present invention is mainly achieved through the following technical solutions: A high-precision pressure sensor compensation method based on deep learning, comprising the following steps: Step S1: Performing temperature tests on the pressure sensor to be tested, the low-precision pressure sensor, and the high-precision pressure sensor shipped from the factory, respectively, to obtain a pressure-temperature output curve to be compensated, a high-noise pressure-temperature output curve, and a standard pressure-temperature output curve; Step S2: training a generator of a generative adversarial network based on the image of the high-noise pressure-temperature output curve and the calibrated pressure-temperature output curve; Step S3: training a discriminator of a generative adversarial network based on the image of the standard pressure-temperature output curve and the calibrated pressure-temperature output curve; Step S4: inputting the image of the pressure-temperature output curve to be compensated into the trained generative adversarial network to obtain an image of the pressure-temperature output curve for preliminary compensation; based on the image of the pressure-temperature output curve for preliminary compensation, converting to obtain preliminary compensated pressure detection values ​​at different temperatures; Step S5: taking the stored historically collected pressure detection values ​​and test temperature values ​​of the pressure sensor to be tested as input, taking the initially compensated pressure detection values ​​at the corresponding temperature as output, training the neural network, and finally obtaining a compensation model for the pressure sensor; Step S6: inputting the detection value of the pressure sensor actually tested and the temperature parameter into the compensation model to obtain the actual compensated pressure detection value.

[0007] In order to better implement the present invention, further, in step S1, based on a set accuracy threshold, a number of other pressure sensors of the same type produced by the same manufacturer are divided into low-precision pressure sensors and high-precision pressure sensors.

[0008] In order to better implement the present invention, further, in step S1, the data in the pressure-temperature output curve to be compensated, the high-noise pressure-temperature output curve and the standard pressure-temperature output curve are normalized respectively, and then the processed pressure-temperature output curve to be compensated, the high-noise pressure-temperature output curve and the standard pressure-temperature output curve are converted into images of the same size to form training sample data.

[0009] In order to better implement the present invention, further, in step S5, a collection cycle is set, and historical data of several collection cycles are stored in the database according to season classification, and the data are stored in different categories based on the range of test temperature values ​​collected within the collection cycle; the neural network is trained based on the stored historical data of the current season.

[0010] In order to better implement the present invention, further, for class storage, if the standard deviation of the test temperature values ​​collected in the current collection period is less than a threshold, the stored historical data is updated, otherwise the test temperature values ​​collected in the current collection period are discarded.

[0011] In order to better implement the present invention, further, in step S5, the neural network is a RBF neural network.

[0012] In order to better implement the present invention, further, the loss function of the compensation model of the pressure sensor is equal to the sum of the loss function of the generative adversarial network and the loss function of the RBF neural network.

[0013] The beneficial effects of the present invention are as follows: The present invention is based on the high-noise pressure-temperature output curves of low-precision pressure sensors, high-precision pressure sensors, and standard pressure-temperature output curves, and trains a high-precision generative adversarial network. Moreover, the training is based on two-dimensional images, and the final network model training result is more accurate. Based on the image of the pressure-temperature output curve to be compensated of the pressure sensor to be tested, the image of the initial compensated pressure-temperature output curve after compensation correction can be obtained through the trained generative adversarial network, and then converted to obtain data of the initial compensated pressure detection value at different temperatures. Finally, the historical data of the pressure sensor to be tested in actual use is combined to train the neural network to obtain the final compensation model of the pressure sensor. The final high-precision pressure detection value is obtained by inputting the pressure detection value and temperature parameters of the online test into the compensation model.

[0014] The present invention can obtain an image of the pressure-temperature output curve of preliminary compensation at the beginning of delivery. In actual use, the user only needs to train the neural network based on the measured historical data and the image of the pressure-temperature output curve of preliminary compensation given by the factory to obtain an accurate compensation model of the pressure sensor, thereby omitting the self-calibration of the pressure, quickly training the model and reducing the capacity of the training data, and has good practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flow chart of a high-precision pressure sensor compensation method based on deep learning in the present invention. DETAILED DESCRIPTION

[0016] Embodiment 1: A high-precision pressure sensor compensation method based on deep learning, such as Figure 1 As shown, the following steps are included: ① Before leaving the factory, temperature tests are performed on piezoresistive pressure sensors of different accuracies produced in the same batch, that is, the pressure source is constant, and by changing the temperature, the pressure sensor outputs different pressure detection values, and then the pressure-temperature output curve is plotted.

[0017] ② Based on the accuracy threshold of the pressure sensor, the pressure-temperature output curve obtained by the test is divided into a high-noise pressure-temperature output curve and a standard pressure-temperature output curve. The data in the high-noise pressure-temperature output curve and the standard pressure-temperature output curve are normalized, and then the processed high-noise pressure-temperature output curve and the standard pressure-temperature output curve are converted into images of the same size, and the images of the calibrated pressure-temperature output curves are combined to form training sample data.

[0018] ③ Based on the images of the high-noise pressure-temperature output curve and the calibrated pressure-temperature output curve, the generator of the generative adversarial network is trained. Then, based on the images of the standard pressure-temperature output curve and the calibrated pressure-temperature output curve, the discriminator of the generative adversarial network is trained.

[0019] ④ Then, the pressure-temperature output curve to be compensated obtained by the temperature test of the pressure sensor to be tested is input into the trained generative adversarial network to obtain an image of the preliminary compensated pressure-temperature output curve; based on the image of the preliminary compensated pressure-temperature output curve, the preliminary compensated pressure detection values ​​at different temperatures are converted.

[0020] ⑤ Before leaving the factory, the image of the preliminary compensated pressure-temperature output curve and the corresponding preliminary compensated pressure detection values ​​at different temperatures are provided to the user.

[0021] ⑥ During actual use, the user continuously collects the pressure detection value and test temperature value of the pressure sensor to be tested and stores the historical data. Then, the historical data is used as input, and the initial compensated pressure detection value at the corresponding temperature is used as output to train the RBF neural network, and finally the compensation model of the pressure sensor is obtained.

[0022] ⑦ Input the detection value of the pressure sensor and the temperature parameters of the actual test into the compensation model to obtain the actual compensated pressure detection value.

[0023] Preferably, in step ⑥, a collection cycle is set, and the historical data of several collection cycles are stored in the database according to the season classification. The data is stored in different categories based on the range of the test temperature values ​​collected in the collection cycle; the neural network is trained based on the stored historical data of the current season. For the category storage, if the standard deviation of the test temperature values ​​collected in the current collection cycle is less than the threshold, the stored historical data is updated, otherwise the test temperature values ​​collected in the current collection cycle are abandoned.

[0024] Specifically, unified training of the generative adversarial network and the RBF neural network can be performed within the manufacturer, or the user and the manufacturer can jointly perform unified training. In this case, the compensation model includes the generative adversarial network and the RBF neural network, and the loss function of the compensation model of the pressure sensor is equal to the sum of the loss function of the generative adversarial network and the loss function of the RBF neural network.

[0025] The present invention can obtain an image of the pressure-temperature output curve of preliminary compensation at the beginning of delivery. In actual use, the user only needs to train the neural network based on the measured historical data and the image of the pressure-temperature output curve of preliminary compensation given by the factory to obtain an accurate compensation model of the pressure sensor, thereby omitting the self-calibration of the pressure, quickly training the model and reducing the capacity of the training data, and has good practicality.

[0026] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification or equivalent change made to the above embodiment based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A high-precision pressure sensor compensation method based on deep learning, characterized in that: The following steps are involved: Step S1: Performing temperature tests on the pressure sensor to be tested, the low-precision pressure sensor, and the high-precision pressure sensor shipped from the factory, respectively, to obtain a pressure-temperature output curve to be compensated, a high-noise pressure-temperature output curve, and a standard pressure-temperature output curve; Step S2: training a generator of a generative adversarial network based on the image of the high-noise pressure-temperature output curve and the calibrated pressure-temperature output curve; Step S3: training a discriminator of a generative adversarial network based on the image of the standard pressure-temperature output curve and the calibrated pressure-temperature output curve; Step S4: inputting the image of the pressure-temperature output curve to be compensated into the trained generative adversarial network to obtain the image of the pressure-temperature output curve for preliminary compensation; Based on the image of the preliminary compensated pressure-temperature output curve, the preliminary compensated pressure detection values ​​at different temperatures are converted; Step S5: taking the stored historically collected pressure detection values ​​and test temperature values ​​of the pressure sensor to be tested as input, taking the initially compensated pressure detection values ​​at the corresponding temperature as output, training the neural network, and finally obtaining a compensation model for the pressure sensor; Step S6: inputting the detection value of the pressure sensor actually tested and the temperature parameter into the compensation model to obtain the actual compensated pressure detection value.

2. A high-precision pressure sensor compensation method based on deep learning according to claim 1, characterized in that: In step S1 , based on a set accuracy threshold, a number of other pressure sensors of the same type produced by the same manufacturer are divided into low-precision pressure sensors and high-precision pressure sensors.

3. The high-precision pressure sensor compensation method based on deep learning according to claim 2 is characterized in that: In step S1, the data in the pressure-temperature output curve to be compensated, the high-noise pressure-temperature output curve and the standard pressure-temperature output curve are normalized respectively, and then the processed pressure-temperature output curve to be compensated, the high-noise pressure-temperature output curve and the standard pressure-temperature output curve are converted into images of the same size to form training sample data.

4. The high-precision pressure sensor compensation method based on deep learning according to claim 1, characterized in that: In step S5, a collection cycle is set, and historical data of several collection cycles are stored in the database according to season classification, and the data are stored in different categories based on the range of test temperature values ​​collected in the collection cycle; The neural network is trained based on the stored historical data of the current season.

5. The high-precision pressure sensor compensation method based on deep learning according to claim 4 is characterized in that: For class storage, if the standard deviation of the test temperature values ​​collected in the current collection period is less than the threshold, the stored historical data is updated, otherwise the test temperature values ​​collected in the current collection period are discarded.

6. The high-precision pressure sensor compensation method based on deep learning according to claim 4, characterized in that: In step S5, the neural network is a RBF neural network.

7. The high-precision pressure sensor compensation method based on deep learning according to claim 6, characterized in that: The loss function of the compensation model of the pressure sensor is equal to the sum of the loss function of the generative adversarial network and the loss function of the RBF neural network.

Images